Universal Dependencies v2: An Evergrowing Multilingual Treebank Collection

In plain words

Universal Dependencies (UD) is a shared rulebook for marking up sentences in many languages, so that one annotated "treebank" can be compared with another. Each word gets a base form, a coarse part-of-speech label, a set of grammar features (such as number or tense), and a link to the word it depends on, with a label saying how. The same labels are meant to mean the same thing in every language, even when the languages use word order, function words or endings differently.

This paper is the second version of that rulebook. It explains what changed from version 1, such as how coordination, ellipsis and multiword expressions are linked, and which new feature values were added. It also lists the collection as it stood in release 2.5: a large set of treebanks covering many languages, some very small and some large. It is a guide and a catalogue, not a new parser or model.

TL;DR

• UD v2 is a cross-lingual annotation scheme for words, morphology and dependency syntax, described as guidelines §2 §3.
• Main v1 to v2 changes: a new oblique relation, a revised coordination attachment, a new flat relation for names and dates, and more features and values §3 tab. 2.
• Release 2.5 has 157 treebanks in 90 languages §4 tab. 3.
• No model, no training and no benchmark: the numbers are corpus statistics reported by the authors.

Problem

Treebanks built by different groups use different labels for the same construction, so parsers and typological studies cannot easily compare or combine them. Earlier schemes (Stanford dependencies, Google POS tags and Interset, as cited in §1) were the starting point; the paper sets out a shared inventory of categories and guidelines so that similar constructions are annotated alike across languages, while letting each language add subtypes §1 §2.

The paper does not state a quantitative failure of the v1 scheme. Its motivation is qualitative (consistency across languages, alignment with the UniMorph project for features §3.2).

Method

The "method" is an annotation scheme with three layers, applied per word.

Word segmentation §2.1. The unit is the syntactic word, not the orthographic token. Clitics and contractions are split into multiword tokens, and in rare cases several tokens are merged into one word.

Morphology §2.2 tab. 1.

1. A lemma: the canonical dictionary form.
2. A universal POS tag from a fixed list of 17 tags. Languages may use a subset but cannot add tags.
3. A set of universal features with name-value pairs. Languages may add language-specific features.

Syntax §2.3 tab. 1 fig. 3. The basic representation is a tree rooted in one word, normally the main predicate. Every word has one head and a label from a fixed inventory of 37 universal relation types. Relations between predicates and their arguments or modifiers are preferred over relations that go through function words §2.3. Relation subtypes such as nsubj:pass and aux:pass mark passive constructions in a language §2.3 fig. 3.

Changes in v2 §3.

• Word-internal spaces: allowed for languages whose writing system marks syllables with spaces, and for a restricted list of exceptions (numbers, abbreviations) that must be listed in the language documentation §3.1.
• Features: new features and values, and some renames, summarised in tab. 2 §3.2.
• Tags: CONJ renamed CCONJ; AUX extended to copulas and non-verbal TAME particles; PART limited to listed words; PRON and DET distinction loosened §3.2.
• Subjects: passive subjects use nsubj (with nsubj:pass recommended); nsubjpass, csubjpass and auxpass removed or subsumed §3.3.
• Oblique nominals at clause level use obl; nmod is reserved for nominal modifiers §3.3.
• cop restricted to pure linking words; neg removed in favour of a Polarity feature; clf added §3.3.
• Coordination: conj still links every non-first conjunct to the first one; cc and punct inside a coordination now attach to the immediately following conjunct, not the first §3.3.
• Ellipsis: a dependent of the elided element is promoted to head, and the remaining siblings attach with orphan, which marks an anomalous structure; the promotion follows an obliqueness hierarchy §3.3.
• Multiword expressions: compound for lexical compounds and phrasal verbs; fixed (formerly mwe) for grammaticalised expressions; flat for headless names, dates and calques. Fixed and flat items attach to the first word as head §3.3.1.

# Attach the parts of a coordination  [§3.3]
# Input:  conjuncts in surface order; markers are the cc and punct tokens
# Output: list of (head, relation, dependent) arcs

def attach_coordination(conjuncts, markers):
    # Every non-first conjunct depends on the first one via conj  [§3.3]
    arcs = []
    for conjunct in conjuncts[1:]:
        arcs.append((conjuncts[0], "conj", conjunct))
    # cc and punct attach to the conjunct that immediately follows them  [§3.3]
    for marker in markers:
        target = next_conjunct_after(marker, conjuncts)
        arcs.append((target, marker.relation, marker))
    return arcs

Enhanced dependencies §3.4. An optional second graph that adds relations and null nodes to make implicit relations explicit. The paper names five enhancement types: null nodes for elided predicates, propagation of shared dependents across conjuncts, controlled and raised subjects, relative pronouns (ref), and case-augmented modifiers such as nmod:on §3.4. Each enhancement is optional §3.4.

Treebank collection §4. The release is described by language, family and genre counts, not by a new data-processing method tab. 3 tab. 4 tab. 5.

Hyperparameters

Not applicable. The paper trains no model and specifies no hyperparameters §1–§5.

Results

Author-reported. This is a resource description: the figures are counts of the UD v2.5 collection as of the paper, with no baseline or model comparison.

Figures: fig. 1 plots languages, treebanks and words from v1.0 to v2.5 as growth curves. fig. 2 plots data size against labelled attachment score for languages in both CoNLL 2017 and CoNLL 2018 shared tasks, comparing the 2017 baseline with the best 2018 result. fig. 4 is a world map of language coverage by release.

Paper vs code

No official code. The paper is a guidelines and resource description and releases no model code. Its links point to universaldependencies.org and lindat.cz §1 §2 §3 §3.4. Searches for the official project repository found no match: the GitHub search on the title terms returned only one unrelated personal repository, and the universaldependencies.org page could not be fetched in this run (web access was not granted), so its links were not checked.

Pitfalls

• Guidelines version vs release number. "UD v2" names the guidelines; "v2.5" names the data release. The text uses both. An implementation must state which one it follows.
• Printed typo in tab. 3. The Kurmanji word count is printed as "1,0260". The paper does not give a corrected value, so do not infer one.
• Language count changes over time. §1 says "approaching 100" languages when written. §4 and the abstract say 90 for release 2.5. Use the release-specific number.
• Family count includes non-families. The 20 families in tab. 4 include Code-Switching, Creole and Sign Language. Germanic (10), Romance (8) and Slavic (13) counts appear only in the text, not in the table §4.
• Removed features are not listed. §3.2 says some features and values were removed, but tab. 2 lists only renamed and added items. A converter cannot take the removals from this paper.
• Coordination changed. A v1 to v2 converter must move cc and punct from the first conjunct to the following conjunct §3.3.
• Relation renames. nsubjpass, csubjpass, auxpass, neg, name, foreign, remnant and mwe change in v2 §3.3 §3.3.1. Old treebanks need mapping.
• Word-internal spaces need a list. Spaces are allowed only in the listed exceptions for non-syllabic languages, and each exception must be declared in the language documentation §3.1. The list itself is not in this paper.
• Enhanced graphs are optional and sparse. Only 24 treebanks in v2.5 had them, and few implemented all five enhancements §3.4.
• Authors' claim about fewer non-projective arcs for the gapping analysis is not backed by a figure in this paper §3.3.

Lineage

• Builds on: UD v1, described in an earlier paper that this text cites as nivre16lrec §1. That paper is not in this wiki. Other sources the text draws on: Stanford dependencies (CCprocessed) §2.3 §3.4, enhanced dependencies §3.4, Google universal POS tags and Interset §1, UniMorph for features §3.2, and the Sag et al. fixed-expression classes §3.3.1.
• Compared against: the v1 guidelines §3 (the paper's own baseline).
• Superseded by: not stated in this paper.